A Linguistically-based Approach to Discourse

Relations Recognition

Rodolfo Delmonte, Gabriel Nicolae, Sanda Harabagiu and Cristina Nicolae

Human Language Technology Research Institute

Department of Computer Science

The University of Texas at Dallas, Richardson, TX 75083

Abstract. We present an unsupervised linguistically-based approach to dis-

course relations recognition, which uses publicly available resources like manu-

ally annotated corpora (Discourse Graphbank, Penn Discourse Treebank, RST-

DT), as well as empirically derived data from “causally” annotated lexica like

LCS, to produce a rule-based algorithm. In our approach we use the subdivision

of Discourse Relations into four subsets – CONTRAST, CAUSE, CONDI-

TION, ELABORATION, proposed by [7] in their paper, where they report re-

sults obtained with a machine-learning approach from a similar experiment,

against which we compare our results. Our approach is fully symbolic and is

partially derived from the system called GETARUNS, for text understanding,

adapted to a specific task: recognition of Causality Relations in free text. We

show that in order to achieve better accuracy, both in the general task and in the

specific one, semantic information needs to be used besides syntactic structural

information. Our approach outperforms results reported in previous papers [9].

1 Introduction

In this paper we will address the task related to discourse processing, i.e. that of Dis-

course Relations Recognition (hence DRR), as well as the other related task, the one

of building Discourse Structures. We take DRR to be of the outmost interest in all

fields of text processing, from summarization to question-answering. We will then

focus on the detection of a particular Discourse Relation, i.e. Causal Relations. The

task referred to as the Detection of Causality Relations (hence DCR) is a precondition

for the possibility that current Q/A systems may advance from factoid based perform-

ances to the answering of WHY questions.

In their paper, [9] report their experiment of Discourse Segmentation into Dis-

course Units in which they use the output of Charniak’s parser in one experiment and

syntactic trees taken directly from the Penn Treebank in another. The recall disagree-

ment is very small: 1.8%, as is the precision: 0.6%, something that speaks in favor of

using syntactic output in a systematic way in the first place. We will return to the

problem of segmentation in a section below. In our approach, Discourse Units will be

segmented automatically, but no learning phase is forecast and the segmentation will

be produced solely on the basis of syntactic structures. This is motivated by the fact

that Discourse Units must be mapped onto semantically interpretable spans of text,

and syntax provides the most natural and reliable interface to semantics available

nowadays (but see [2]). The problem is addressed in more detail in the section on

Evaluation below.

Delmonte R., Nicolae G., Harabagiu S. and Nicolae C. (2007).

A Linguistically-based Approach to Discourse Relations Recognition.

In Proceedings of the 4th International Workshop on Natural Language Processing and Cognitive Science, pages 81-91

DOI: 10.5220/0002421300810091

 SciTePress

The learning algorithm presented both in [7] and [9] makes no use of semantic in-

formation, but only of lexical, grammatical and syntactic information. In particular, in

the former the classifier makes use of what they call the “most representative words”

of a sentence/discourse unit, by which they mean the nouns, verbs and cue phrases

present in each sentence. This is just a crude erasing procedure of function words,

including all adjectives and adverbs. On the contrary, in [9] the training section for

the segmentation task consisted of triples that contained the syntactic tree and the dis-

course tree.

The discourse model the authors use for Discourse Parsing, on the contrary, con-

tains information directly related to what they call “a lexicalized syntactic tree”, very

much in line with Collins’ approach to parsing. In their model, each feature is a repre-

sentation of the syntactic and lexical information that is found at the point where two

EDUs are joined together. For each EDU they thus identify a word as the Head word,

denoted H, which is the highest node in EDU. Thus, the verb will always be the Head

of finite tensed EDUs and, in case a modal or an auxiliary precedes the main verb, it

will be taken as Head, since it will be mapped onto a higher VP node than the main

verb. This would be semantically unacceptable in our approach, where the syntax

only plays a secondary role.

We carried out a thorough study of the three databases – Discourse Graphbank,

hence DGB

[10], Penn Discourse Treebank, hence PDTB

[8], and RST- DT

[1] -

which contain manually annotated Discourse Relations, with the aim to propose a

Model of Causal Relations (hence CRs) that encompasses different levels of linguistic

representations and features. The Model will eventually be used by a ques-

tion/answering system to enable it to deal with WHY questions. CRs have always

constituted a very thorny issue in the linguistic and philosophical literature, and to

cover all related issues we will address the whole gamut of linguistic information.

The paper is organized as follows: in section 2 we introduce the theoretical

framework, the data analysis for the DCRs and the difference between explicit and

implicit discourse relations; section 3 presents the Hypothesis and the Model; section

4 is devoted to the evaluation, and finally there is a short conclusion.

2 Causal Relations

The goal of this work was finding if and in what measure CRs are dependent on

Knowledge of the World, and what kind of constraints surface syntactic structure and

semantic interpretation play on the nature of CRs— eventually what kind of informa-

tion needs to be stored in the lexicon in order to spot CRs.

As a general starting point, we noted that natural CRs – i.e. causality relations re-

lated to natural facts of the physical world in which we live - are usually left unex-

pressed unless the text is dealing with explanatory descriptions and experimental

work in scientific literature and educational academic and school related literature.

Whenever CRs are reported of physical events they must be exceptional or possibly of

http://www.ldc.upenn.edu/Catalog/CatalogEntry.jsp?catalogId=LDC2005T08

http://www.seas.upenn.edu/~pdtb

http://www.ldc.upenn.edu/Catalog/CatalogEntry.jsp?catalogId=LDC2002T07

abnormal dimension, e.g., the tsunami is an abnormal ocean wave which is caused by

landslide or earthquake.

In order to express CRs in text, two clauses/sentences must be used: one express-

ing Motivation or Cause, and another expressing Result or Effect. It is not always the

case that Causes are expressed fully: in many cases they are left implicit as if they can

be understood by the addressee. What is usually reported in texts is the effect of some

previous event which might or might not be reported in the previous text, and may be

adjacent or (rarely) non adjacent to the current sentence. We also noted that in DGB

CRs expressed at a propositional level, thus intersententially, make use of a connec-

tive only in 30% of cases: the remaining examples are either untensed propositions –

gerundives, infinitivals or adjuncts – and don’t carry any specific or unambiguous

discourse marker, or they are sentences without the presence of a connective, thus

leaving the link implicit. The latter constitute certainly the most challenging type of

CRs to detect automatically. We have also considered situations in which Causes are

not expressed by clauses, but simply by Prepositional Phrases, as for instance in the

example below that we take from DGB:

{The late Prime Minister Olof Palme was blacklisted from the White House}

effect

{for his

outspoken opposition to the Vietnam War.}

cause

Eventually, Cause may be simply part of the lexical meaning of the verb/noun,

and may be extracted by some system of semantic decomposition in lexical primitives

or templates. This is done using the LCS lexicon

, where Causatives and Inchoatives

are clearly marked by the presence of a CAUSE template.

In order to isolate difficult to solve cases from the rest, we produced a first subdi-

vision of possibly different types of CRs along the lines proposed by a number of pre-

vious works, and we came up with the following basic inventory that we adopted in

our algorithm:

- Cause-Result

- Rationale-Outcome

- Purpose-Outcome

- Circumstance-Outcome

- Means-Outcome

We use the label Outcome to indicate an Event which is not in direct causal rela-

tion with another event but is simply explained (Rationale) by it, marked with the

intention to carry it out (Purpose), or in a possible or conditional relation with it (Cir-

cumstance). We will treat other dubious cases as the ones indicated below as Circum-

stances:

Vitamin B enables the body to digest food.

Corn oil prevents butter from burning.

The reason for this choice is that cases like these contain Generic assertions and

not Factual assertions, and we want to keep Factual assertions apart from Nonfactual

ones. In this sense, we also consider conditionals and counterfactuals as belonging to

a totally different class of Discourse Relations, namely the one definable as Hypothe-

sis. Eventually, temporal discourse markers such as after, when will not be treated as

http://www.umiacs.umd.edu/~bonnie/verbs-English.lcs

possible causal cue words, but just as temporal ones or as Circumstances in case they

imply some causal link, as for instance in the following two examples:

Liberman left the Republican Party after George W. Bush was nominated.

Mary stopped the car when she saw the elephant.

2.1 Data Analysis: The CRs Examples Annotated in DGB and PDTB

The data analysis has been concerned with the individuation of linguistic criteria to

subdivide CRs as they have been selected and annotated in the two databanks avail-

able at HLT. DGB is the corpus of annotated Discourse Relations built by [10], which

contains 8910 relations and clause pairs, 466 (thus constituting approximately the 5%

of all relations) of which have been labeled as CE, i.e. Cause-Effect. The PDTB con-

tains only 1515 discourse relations, 204 of which have been labeled as CAUSE (ap-

proximately 13%).

The two corpora are markedly different however: the PDTB only contains “im-

plicit” or lexically unexpressed discourse markers. In other words, sentence pairs re-

lated by a given Discourse Relation are not connected by an explicit lexical connec-

tor. They are also usually made of two separate sentences; however, the structures

selected may also respond to phonological criteria, as reported in the Technical Man-

ual, and in some cases they may be constituted by PPs that are adjuncts contained in a

parenthetical. In such cases, punctuation plays a role, together with intonation. The

same applies to the other database we inspected.

As far as the DGB is concerned, discourse relations may come with both ex-

pressed and unexpressed connectors, both inter- and intrasententially organized. In

particular, as far as CRs are concerned, we have the following distribution of data:

- 190 sentence pairs: 1) lexically unexpressed discourse marker;

- 136 clause pairs: 1) result/effect expressed as infinitival preceded by

to/in_order_to;

- 89 sentence pairs: 1) lexically expressed discourse markers (68 because + 21

because_of + 4 cause, caused_by);

- 38 clause pairs: 1) result/effect expressed as gerundive preceded by by/for; 2)

intrasententially;

- 35 sentence pairs: 1) lexically expressed discourse markers (as, as_a_result, so.

so_that, as_a_means_of, part_of, since, since_then, when, due_to);

- 8 have not been classified because we were not in agreement with the annota-

tors.

Overall, we ended up with 2/3 of all CRs with no specific lexical connectors and

1/3 having one specific discourse marker, 2/5 of all CRs constituted by intrasentential

structures and 3/4 having an intersentential structure. Thus, by far, the majority of

discourse segments annotated are not between two clauses with a finite verb. Other

additional semantic and syntactic details we noted are:

- 70 contain negation markers (not, no, none, nothing, never, neither, nor);

- 39 sentences treated as DUs are Relative Clauses;

- 82 NPs intervene in the discourse relations either directly as one segment or

indirectly;

- 13 PPs and APs are treated as DU segments.

As to PDTB, we analyzed all of the 204 CRs pairs and we discovered that they are

highly paraphrastic, using copulative or light verbs together with extended nominali-

zation to convey the meaning of events. Overall, we counted 218 cases of copulative

verbs as main verbs (be, there_be, become, have), which makes at least one for each

sentence pair. On the contrary, the same verbs in the DGB only amount to 126 cases,

only about 27% of the sentence pairs. We also counted the number of clauses overall,

and the mean length per discourse relation is much higher in PDTB (3.1), when com-

pared to DGB (2.3).

In PDTB, Cause annotated sentence pairs show a consistent use of negation: there

are 83 sentences containing negation, i.e. 40.7%; if we compare this with DGB, where

we had 70 negation items – i.e. 15% of the all examples – we can see that it is much

higher.

2.2 Explicit and Implicit CRs: Are they Interchangeable?

For computational purposes, we wanted to assume that the use of discourse markers is

not optional. The strong hypothesis would be that in order to be able to omit/insert a

causality discourse marker some structural and semantic conditions have to be ob-

tained. In doing this work we discovered that inserting BECAUSE in implicit CR-

DUs corresponds to testing the causality reading of the DR. Consider the following

examples taken from the DGB:

DU1: to produce chemical weapons.

DU2: the plant was built

- The plant was built BECAUSE they/one wanted/intended to produce chemical weapons.

DU1: He died in Dec. 21 Pan Am plane crash in Scotland while on his way to New York

DU2: to observe the signing of an agreement for Namibian independence from South Afri-

can rule.

- He died in Dec. 21 Pan Am plane crash in Scotland while on his way to New York BE-

CAUSE he wanted/intended to observe the signing of an agreement for Namibian inde-

pendence from South African rule.

However, to enforce our starting hypothesis of the non comparability of implicitly

vs. explicitly marked CRs, no such reformulation is allowed when the relation is actu-

ally a CAUSE DR, i.e. there are two separate clauses with finite tense, and there is no

discourse marker. Consider:

DU1: these early PCs triggered explosive product development in desktop models for the

home and office.

DU2: PC shipments annually total some $38.3 billion world - wide.

- These early PCs triggered explosive product development in desktop models for the home

and office BECAUSE PC shipments annually total some $38.3 billion world-wide.

So, eventually, we may safely affirm that untensed clauses cannot be classified as

Causal Relations in case there is no introductory preposition – typical ones are: to,

in_order_to, thanks_to, for, from, by. As to tensed clauses, the examples with no cue

word are markedly different and contain special assertions, heavily paraphrastic and

idiomatic. This justifies a model based on a bipartition of clauses, where finite ones

will only convey Causal Relations in exceptional cases and only in presence of a se-

mantic marker.

3 The Hypothesis

We assumed as our starting hypothesis that CAUSES must be expressed whenever the

EFFECTS are unexpected, unwanted, unattested, in other words whenever the FACT

constituting the effect clause is new, surprising, negates or contradicts the evidence.

To achieve such a pragmatic result, EFFECT clauses must contain at least one of the

major or more than one of the following linguistic components:

A. Minor components:

- NEGATION

- MODALITY OPERATOR (adverb, modal verb, adjective, etc.)

- QUANTITY OPERATOR (quantifiers, intensifiers, adjectives, etc.)

- DISCOURSE MARKERs indicating ADVERSE or CONTRARY meaning (but).

B. Major components:

- DISCOURSE MARKERs indicating cause (because_of, caused_by, due_to, as,

etc.)

- DISCOURSE MARKERs indicating result (why, result_from, as_a_result,

so_that, so)

- LEXICALIZED NEGATION (refuse, reject)

- ATTITUDINAL VERBS (think, believe, expect, criticize, attack, etc.)

- EMOTIONAL VERBS (concern, worry, doubt, fear, etc.)

- QUANTITY GRADING VERBS (rise, soar, slump, cut_down, etc)

- EVALUATIVE PREDICATES (wrong, right, improve, worse, etc.)

- NECESSITY PREDICATES (must, ought_to, need, have_to, etc.)

- EVENTIVE DEVERBAL NOUNS and ADJECTIVES indicating (natural) catas-

trophe, illegal or dangerous events (attack, crash, weapon, bribe, etc.)

- PREDICATES (adjectives, adverbials, verbs, nouns) indicating novelty, discovery

(scientific and not)

- PREDICATES (adjectives, adverbials, verbs, nouns) indicating problematic and

troublesome situations and events.

C. Structural components:

- RESULTATIVE INFINITIVALS

- ADJUNCT GERUNDIVES headed with for, by

- ADJUNCT PPs OR NPs whose nominal head is deverbal or indicates a state-result

in aspectual terms.

D. Referential components:

- SUBJect pronominals corefer

- Possessive pronominal corefers to a previous argument

- The second sentence in the discourse pair starts with a deictic pronoun (this/that).

3.1 Towards a General Model of CRs

Attempts at producing Automatic Discourse Relations Classification are very few in

the literature: Ken Barter in 1995 produced an interactive algorithm and an evalua-

tion, limited though to a small number of examples. As reported above, [7] carried out

a study with the aim of detecting discourse relations automatically. They trained an

automatic classifier to recognize the relations that were not signaled by a discourse

marker. The classifier was trained on examples of actual relations versus examples of

non-relations (random pairs of units, sometimes each taken from different docu-

ments), using lexical patterns. It learnt to distinguish relations that were not signaled

by a discourse marker, increasing accuracy over a discourse-marker-based method by

as much as 77%. This result refers to an experiment carried out on automatically clas-

sified DRs. When tested on manually annotated data, the same algorithm has been

reported to achieve an accuracy of 57%.

We now comment [9]’s paper, where the authors report on another experiment in-

tended to improve on the results reported in [6]. The experiment is organized around

two tasks: segmenting the text into Discourse Units, on the basis of DUs organized

with lexical heads and syntactic constituency, and producing an automatic labeling of

the test data. The final results fare better that the previous ones; F-measure is around

68%.

The model we are thinking of is illustrated in Figure 1 below, and is based on

three main sources of information:

- structural information, to tell CAUSE apart from the remaining relations;

- lexical information, to detect complementation cases of complement nontensed

clauses – infinitivals and gerundives – from ADJUNCT, which alone can con-

tribute to NON-CAUSE relations;

- tense and reference as constraints on the definition of CAUSE relations.

Fig. 1. A scheme of the algorithm to detect CRs with symbolic procedures.

In addition to these main factors, we will use knowledge of the world in the form of

evaluative, subjective, Exceptional Event predicates, predicates of negative judgment,

etc. as they have been defined in the Hypothesis section above.

The system uses the output of a parser – in our case GETARUNS’ parser – made

up of Functional Dependency relations, where we added a number a semantic items to

complete the interpretation task. At first the clauses are subdivided into two structural

types: tensed clauses are set apart from untensed clauses. Untensed clauses are

checked for Non-Argumenthood before they may be considered as DUs. They con-

tribute a Result interpretation that may trigger the Cause relation in the main clause.

Else, they are interpreted as Rationale or Purpose and trigger an Outcome relation in

the main clause. Gerundives preceded by preposition by will be interpreted as Means;

those preceded by for are interpreted as Rationale or Purpose according to additional

conditions obtained in the main clause – see below.

As to tensed clauses, be they relative clauses, complement clauses, coordinate

clauses, or simply main clauses, they may express all gamut of Causal Relations. The

most difficult to be assigned, as we noted above, is Circumstance. It requires a certain

number of constraints to apply before the interpretation is accepted by the algorithm.

In particular, the SUBJect argument must be non-animate non-human, or else the

proposition must be under the scope of modal operators, or the main predicate must

be a semantically opaque predicate. In the presence of subordinators, specific cue

words or causality markers, the discourse relation is assigned straightforwardly. After

that, the choice is still difficult to make: the algorithm will use the list of Exceptional

Event predicates, Negative Judgment predicates, as well as the presence of negation

or other indicators of some uncommon or unpredictable situation that requires the

Cause to be explicitly expressed. Finally, if some of these are obtained, the algorithm

looks for cohesive links by accessing the output of the anaphora resolution algorithm

and checks for the presence of Resolved Nominals or Pronominals, i.e. whether the

current clause contains nouns/pronouns which corefer to some previously mentioned

entity present in the history list. In case this also fails and there is no cohesive link,

the system will have to search WordNet, LCS and FrameNet for some indication that

the Predicate-Argument Structures contained in the two clauses under examination

can be taken to be explicitly in a Causality Relation. We usually start out by a lookup

in LCS. LCS entries contain cross reference to Levin verb classes, to WordNet senses

and to PropBank argument lists. These have been mapped to a more explicit label set

of Semantic Roles, which can be regarded more linguistically motivated than the ones

contained in FrameNet, which are more pragmatically motivated. However, for our

purposes, LCS notation is more perspicuous because of the presence of the CAUSE

operator, and is more general: from a total number of 9000 lexical entries, 5000 con-

tain the CAUSE operator. On the contrary, in FrameNet, from a total number of

10000 lexical entries, only 333 are related to a CAUSE Frame; if we search for the

word “cause” in the definitions contained in all the Frames, the number increases to

789, but is still too small compared to LCS.

3.2 The CRs Algorithm in Detail

We will now present in detail the contents of the algorithm from a technical and lin-

guistic point of view. The algorithm performs these actions:

1. collects linguistic information from the parser output;

2. translates linguistic information into Semantic and Pragmatic Classes;

3. assigns Discourse Relations on a clause-by-clause level: some will be spe-

cific, some generic or default;

4. detects the presence of Causality Relations according to the algorithm traced

above;

5. determines coreference relations between the Discourse Units thus classified.

The input to the algorithm is reported in Table 1: the list of linguistic items repre-

sents clause level information as derived from dependency structure – related to the

example “he was not frozen in place by rigid ideology”:

Table 1. Linguistic information from dependency structure.

Speech Predicate Tense Voice Modal

direct freeze past passive nil

Neg Aux Foc Aspect Mood

not nil nil Accomplishment Indic

Lexical aspect is derived from LCS following the approach presented in [5]. The

remaining linguistic items are quite straightforward to describe from the output of a

dependency parser. Table 2 contains the semantic interpretation of the initial features,

which is wrought by means of combining features together and the use of external

knowledge to tell different classes of verbs apart, which we take from WordNet, LCS,

FrameNet and PropBank. For instance, Auxiliary and Tense contribute to produce

Comptense; Point of view depends on Speech and the semantic class of verb; Factual-

ity depends on Mood and Modality; Relevance depends on negation and Viewpoint.

A full set of rules is reported in [4] (to be published).

Table 2. Informational structure derived from linguistic information.

ClauseType/SemRole View-point Comptense Factuality Relevance

main/prop external past no Back-ground

The second step takes as input the forms shown above in Table 2 and the list of

arguments and adjuncts, and produces the output of Table 3. The second step is repre-

sented below by three sets of rules:

1. the first rule looks for clauses governed by be/have, i.e. copulative constructions in

which the verb has no semantic relevance;

2. the second rule looks for all cases of clauses, be it finite or nonfinite;

3. the third rules takes into account adjunct PPs governed by a preposition, which

have already received their interpretation by the parser.

Table 3. Discourse structure derived from informational structure.

DISCREL DISCDOMAIN DISCSTRUCTURE

narration objective down (1-2)

Finally, as explained above and indicated in Figure 1, the rules elaborated for the

detection of Causal Relations are fired after the other DRs have already been decided.

The rules have the task of recovering the Cause/Circumstance whenever a Re-

sult/Effect has already been decided by the previous algorithm. To this aim, we try to

take advantage of the information encoded in two resources: WordNet and FrameNet.

WordNet contains a small list of Cause-Effect inferential links between verb predi-

cates – some 200. FrameNet encodes the same information in a few Frame to Frame

relations, 12 – 14 if we add Inchoatives. We also adapted the information available in

VerbOcean

, which, however, only focuses on temporal restrictions. It has a list of

HAPPENS-BEFORE relations amounting to 6500, but only a small percentage of

those can actually be computed as entertaining some causal relation.

In order to spot Negative Judgment predicates we use the Laswell Value Diction-

ary and the Harvard IV dictionary

. As discussed above, DISCOURSEREL is recom-

puted after the first pass.

http://semantics.isi.edu/ocean/

http://www.webuse.umd.edu:9090/tags/

In case a RESULT or a RATIONAL clause has been previously computed, the

system looks for the most adequate adjacent clause to become a CAUSE or an OUT-

COME. This is done by checking Temporal and Coreferentiality constraints. The out-

come of this pass is to relabel some previously labeled Discourse Relations that con-

stitute default relations into the one required by the context.

4 Results and Evaluation

In order to carry out the evaluation, we assumed that the gap existing between Dis-

course Units (hence DUs) and Syntactic Units (hence SUs) should be minimal. To

that end, we compared the three corpora at hand with the task to be realized. As it

turned out, PDTB resulted unsuitable for both tasks, the DRR and the DSR. The rea-

son for that is simply the arbitrariness of correspondences existing between Syntactic

Units and Discourse Units. The annotators had syntactic structures available from

PTB and did not impose themselves a strict criterion for the definition of what a DU

might correspond to. The result is an average proportion of 3.2 clauses per DU, a fact

that makes the corpus unsuitable for either tasks. In their presentation of annotation

guidelines, downloadable from PDTB website, [8] summarize in one slide what

counts as a legal Discourse Unit, which they refer to as Argument: a single clause

(tensed/untensed); a single sentence; multiple sentences; NPs that refer to clauses; and

some nominal forms expressing events or states. So, basically, any syntactic unit may

actually become a discourse Argument, and the correspondence Clause-DU is com-

pletely lost. The problem is that, in order to take automatic decisions, the criteria to be

used become hard to define, and refer to the pragmatic and semantic domain rather

than simply and more safely to syntax. This situation should also affect machine

learning methods, which would use general linguistic features to build their models

and then would be at a loss when trying to scale them to different syntactic structures.

So, finally, we are left with two corpora, DGB and RST. The latter in particular is

very strictly compliant with a syntax based unit correspondence. DGB, even though

not intended to respond to such a criterion, has an average ratio of clauses per DU of

2.2. We decided to cope with this inconsistency by assigning a Discourse Relation to

a corresponding larger Unit in case the smaller one was missing when computing the

same sentence, on the basis of the presence of the same main predicate.

General data for the two corpora are tabulated below:

Table 4. General data and EDU/sentence ratio.

No. of sentences No. of words No. of EDUs RatioEDU/Ss

RST-test

799 19300 2346 2.9

DGB

3110 72520 8910 2.8

Table 5. Quantitative evaluation of DU segmentation.

No. of EDUs My no. of EDUs Recall Precision

RST-test

2346 2281 97.2 87.9

DGB

8910 8910 100 89.8

Results reported in [9] report an F-measure of 84.7% when using the PennTree-

bank structure and 83.1% when using Charniak’s parser.

Table 6. F-measure values for the DGB experiment.

Causality relations Condition Contrast Other relations

DGB

38.13 48.25 44.34 59.13

The overall F-measure for DGB is 64.57%.

5 Conclusions

We presented an algorithm for the automatic detection of discourse relations, particu-

larly focused on Causality Relations, that works on linguistically-based symbolic

rules and needs no training. The paper was mainly concerned with the work carried

out on characterizing Causality Relations from a syntactic, semantic and pragmatic

point of view. The corpus we used to test our system, Discourse Graphbank, has not

yet been used by other research teams, and there is no comparison we can make.

However, we intend to use it and test it with other corpora. Our algorithm only con-

tains general rules and no special procedure to cope with implicit Discourse Relations.

We may safely assume that once these rules have been implemented and tested exten-

sively on the corpora available, an increase in the accuracy is to be expected.

References

1. Carlson L., Marcu, D. and Okurowski M. 2002. RST Discourse Treebank. Philadelphia,

PA:Linguistic Data Consortium.

2. Delmonte R. 2003. Parsing Spontaneous Speech. In proceedings of EUROSPEECH2003.

3. Delmonte R. 2005. Deep & Shallow Linguistically Based Parsing. In A.M. Di Sciullo (ed.),

UG and External Systems, John Benjamins, Amsterdam/Philadelphia, pp. 335-374.

4. Delmonte R. 2007 (to be published). Computational Linguistic Text Processing. Nova Pub-

lishers, New York.

5. Dorr B. J. and Olsen M. B. 1997. Deriving Verbal and Compositional Lexical Aspect for

NLP Applications. In proceedings of the 35th ACL, pp. 151-158.

6. Marcu, D. 2000. The theory and practice of discourse parsing and summarization. Cam-

bridge, MA: MIT Press.

7. Marcu D. and Echihabi A., 2002. An Unsupervised Approach to Recognizing Discourse

Relations. In proceedings of the 40th Annual Meeting of the Association for Computational

Linguistics (ACL'02), Philadelphia, PA, pp. 368-375.

8. Miltsakaki E., Prasad R., Joshi A. and Webber B. 2004. The Penn Discourse Treebank. In

proceedings of the 4th International Conference on Language Resources and Evaluation,

Lisbon, Portugal.

9. Soricut R. and Marcu D. 2003. Sentence level discourse parsing using syntactic and lexical

information. In proceedings of the 2003 Conference of the North American Chapter of the

Association for Computational Linguistics on Human Language Technology-Volume 1, pp.

149–156.

10. Wolf F., Gibson E., Fisher A., Knight M. 2005. Discourse Graphbank. Linguistic Data

Consortium, Philadelphia.